2 * net/sched/sch_htb.c Hierarchical token bucket, feed tree version
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; either version
7 * 2 of the License, or (at your option) any later version.
9 * Authors: Martin Devera, <devik@cdi.cz>
11 * Credits (in time order) for older HTB versions:
12 * Stef Coene <stef.coene@docum.org>
13 * HTB support at LARTC mailing list
14 * Ondrej Kraus, <krauso@barr.cz>
15 * found missing INIT_QDISC(htb)
16 * Vladimir Smelhaus, Aamer Akhter, Bert Hubert
17 * helped a lot to locate nasty class stall bug
18 * Andi Kleen, Jamal Hadi, Bert Hubert
19 * code review and helpful comments on shaping
20 * Tomasz Wrona, <tw@eter.tym.pl>
21 * created test case so that I was able to fix nasty bug
23 * spotted bug in dequeue code and helped with fix
25 * fixed requeue routine
26 * and many others. thanks.
28 #include <linux/module.h>
29 #include <linux/moduleparam.h>
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/string.h>
33 #include <linux/errno.h>
34 #include <linux/skbuff.h>
35 #include <linux/list.h>
36 #include <linux/compiler.h>
37 #include <linux/rbtree.h>
38 #include <linux/workqueue.h>
39 #include <linux/slab.h>
40 #include <net/netlink.h>
41 #include <net/sch_generic.h>
42 #include <net/pkt_sched.h>
43 #include <net/pkt_cls.h>
47 ========================================================================
48 HTB is like TBF with multiple classes. It is also similar to CBQ because
49 it allows to assign priority to each class in hierarchy.
50 In fact it is another implementation of Floyd's formal sharing.
53 Each class is assigned level. Leaf has ALWAYS level 0 and root
54 classes have level TC_HTB_MAXDEPTH-1. Interior nodes has level
55 one less than their parent.
58 static int htb_hysteresis __read_mostly = 0; /* whether to use mode hysteresis for speedup */
59 #define HTB_VER 0x30011 /* major must be matched with number suplied by TC as version */
61 #if HTB_VER >> 16 != TC_HTB_PROTOVER
62 #error "Mismatched sch_htb.c and pkt_sch.h"
65 /* Module parameter and sysfs export */
66 module_param (htb_hysteresis, int, 0640);
67 MODULE_PARM_DESC(htb_hysteresis, "Hysteresis mode, less CPU load, less accurate");
69 static int htb_rate_est = 0; /* htb classes have a default rate estimator */
70 module_param(htb_rate_est, int, 0640);
71 MODULE_PARM_DESC(htb_rate_est, "setup a default rate estimator (4sec 16sec) for htb classes");
73 /* used internaly to keep status of single class */
75 HTB_CANT_SEND, /* class can't send and can't borrow */
76 HTB_MAY_BORROW, /* class can't send but may borrow */
77 HTB_CAN_SEND /* class can send */
86 /* When class changes from state 1->2 and disconnects from
87 * parent's feed then we lost ptr value and start from the
88 * first child again. Here we store classid of the
89 * last valid ptr (used when ptr is NULL).
94 /* interior & leaf nodes; props specific to leaves are marked L:
95 * To reduce false sharing, place mostly read fields at beginning,
96 * and mostly written ones at the end.
99 struct Qdisc_class_common common;
100 struct psched_ratecfg rate;
101 struct psched_ratecfg ceil;
102 s64 buffer, cbuffer;/* token bucket depth/rate */
103 s64 mbuffer; /* max wait time */
104 u32 prio; /* these two are used only by leaves... */
105 int quantum; /* but stored for parent-to-leaf return */
107 struct tcf_proto __rcu *filter_list; /* class attached filters */
108 struct tcf_block *block;
111 int level; /* our level (see above) */
112 unsigned int children;
113 struct htb_class *parent; /* parent class */
115 struct net_rate_estimator __rcu *rate_est;
118 * Written often fields
120 struct gnet_stats_basic_packed bstats;
121 struct tc_htb_xstats xstats; /* our special stats */
123 /* token bucket parameters */
124 s64 tokens, ctokens;/* current number of tokens */
125 s64 t_c; /* checkpoint time */
128 struct htb_class_leaf {
129 int deficit[TC_HTB_MAXDEPTH];
132 struct htb_class_inner {
133 struct htb_prio clprio[TC_HTB_NUMPRIO];
138 int prio_activity; /* for which prios are we active */
139 enum htb_cmode cmode; /* current mode of the class */
140 struct rb_node pq_node; /* node for event queue */
141 struct rb_node node[TC_HTB_NUMPRIO]; /* node for self or feed tree */
143 unsigned int drops ____cacheline_aligned_in_smp;
144 unsigned int overlimits;
148 struct rb_root wait_pq;
149 struct htb_prio hprio[TC_HTB_NUMPRIO];
153 struct Qdisc_class_hash clhash;
154 int defcls; /* class where unclassified flows go to */
155 int rate2quantum; /* quant = rate / rate2quantum */
157 /* filters for qdisc itself */
158 struct tcf_proto __rcu *filter_list;
159 struct tcf_block *block;
161 #define HTB_WARN_TOOMANYEVENTS 0x1
162 unsigned int warned; /* only one warning */
164 struct work_struct work;
166 /* non shaped skbs; let them go directly thru */
167 struct qdisc_skb_head direct_queue;
170 struct qdisc_watchdog watchdog;
172 s64 now; /* cached dequeue time */
174 /* time of nearest event per level (row) */
175 s64 near_ev_cache[TC_HTB_MAXDEPTH];
177 int row_mask[TC_HTB_MAXDEPTH];
179 struct htb_level hlevel[TC_HTB_MAXDEPTH];
182 /* find class in global hash table using given handle */
183 static inline struct htb_class *htb_find(u32 handle, struct Qdisc *sch)
185 struct htb_sched *q = qdisc_priv(sch);
186 struct Qdisc_class_common *clc;
188 clc = qdisc_class_find(&q->clhash, handle);
191 return container_of(clc, struct htb_class, common);
194 static unsigned long htb_search(struct Qdisc *sch, u32 handle)
196 return (unsigned long)htb_find(handle, sch);
199 * htb_classify - classify a packet into class
201 * It returns NULL if the packet should be dropped or -1 if the packet
202 * should be passed directly thru. In all other cases leaf class is returned.
203 * We allow direct class selection by classid in priority. The we examine
204 * filters in qdisc and in inner nodes (if higher filter points to the inner
205 * node). If we end up with classid MAJOR:0 we enqueue the skb into special
206 * internal fifo (direct). These packets then go directly thru. If we still
207 * have no valid leaf we try to use MAJOR:default leaf. It still unsuccessful
208 * then finish and return direct queue.
210 #define HTB_DIRECT ((struct htb_class *)-1L)
212 static struct htb_class *htb_classify(struct sk_buff *skb, struct Qdisc *sch,
215 struct htb_sched *q = qdisc_priv(sch);
216 struct htb_class *cl;
217 struct tcf_result res;
218 struct tcf_proto *tcf;
221 /* allow to select class by setting skb->priority to valid classid;
222 * note that nfmark can be used too by attaching filter fw with no
225 if (skb->priority == sch->handle)
226 return HTB_DIRECT; /* X:0 (direct flow) selected */
227 cl = htb_find(skb->priority, sch);
231 /* Start with inner filter chain if a non-leaf class is selected */
232 tcf = rcu_dereference_bh(cl->filter_list);
234 tcf = rcu_dereference_bh(q->filter_list);
237 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
238 while (tcf && (result = tcf_classify(skb, tcf, &res, false)) >= 0) {
239 #ifdef CONFIG_NET_CLS_ACT
244 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
250 cl = (void *)res.class;
252 if (res.classid == sch->handle)
253 return HTB_DIRECT; /* X:0 (direct flow) */
254 cl = htb_find(res.classid, sch);
256 break; /* filter selected invalid classid */
259 return cl; /* we hit leaf; return it */
261 /* we have got inner class; apply inner filter chain */
262 tcf = rcu_dereference_bh(cl->filter_list);
264 /* classification failed; try to use default class */
265 cl = htb_find(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
266 if (!cl || cl->level)
267 return HTB_DIRECT; /* bad default .. this is safe bet */
272 * htb_add_to_id_tree - adds class to the round robin list
274 * Routine adds class to the list (actually tree) sorted by classid.
275 * Make sure that class is not already on such list for given prio.
277 static void htb_add_to_id_tree(struct rb_root *root,
278 struct htb_class *cl, int prio)
280 struct rb_node **p = &root->rb_node, *parent = NULL;
285 c = rb_entry(parent, struct htb_class, node[prio]);
287 if (cl->common.classid > c->common.classid)
288 p = &parent->rb_right;
290 p = &parent->rb_left;
292 rb_link_node(&cl->node[prio], parent, p);
293 rb_insert_color(&cl->node[prio], root);
297 * htb_add_to_wait_tree - adds class to the event queue with delay
299 * The class is added to priority event queue to indicate that class will
300 * change its mode in cl->pq_key microseconds. Make sure that class is not
301 * already in the queue.
303 static void htb_add_to_wait_tree(struct htb_sched *q,
304 struct htb_class *cl, s64 delay)
306 struct rb_node **p = &q->hlevel[cl->level].wait_pq.rb_node, *parent = NULL;
308 cl->pq_key = q->now + delay;
309 if (cl->pq_key == q->now)
312 /* update the nearest event cache */
313 if (q->near_ev_cache[cl->level] > cl->pq_key)
314 q->near_ev_cache[cl->level] = cl->pq_key;
319 c = rb_entry(parent, struct htb_class, pq_node);
320 if (cl->pq_key >= c->pq_key)
321 p = &parent->rb_right;
323 p = &parent->rb_left;
325 rb_link_node(&cl->pq_node, parent, p);
326 rb_insert_color(&cl->pq_node, &q->hlevel[cl->level].wait_pq);
330 * htb_next_rb_node - finds next node in binary tree
332 * When we are past last key we return NULL.
333 * Average complexity is 2 steps per call.
335 static inline void htb_next_rb_node(struct rb_node **n)
341 * htb_add_class_to_row - add class to its row
343 * The class is added to row at priorities marked in mask.
344 * It does nothing if mask == 0.
346 static inline void htb_add_class_to_row(struct htb_sched *q,
347 struct htb_class *cl, int mask)
349 q->row_mask[cl->level] |= mask;
351 int prio = ffz(~mask);
352 mask &= ~(1 << prio);
353 htb_add_to_id_tree(&q->hlevel[cl->level].hprio[prio].row, cl, prio);
357 /* If this triggers, it is a bug in this code, but it need not be fatal */
358 static void htb_safe_rb_erase(struct rb_node *rb, struct rb_root *root)
360 if (RB_EMPTY_NODE(rb)) {
370 * htb_remove_class_from_row - removes class from its row
372 * The class is removed from row at priorities marked in mask.
373 * It does nothing if mask == 0.
375 static inline void htb_remove_class_from_row(struct htb_sched *q,
376 struct htb_class *cl, int mask)
379 struct htb_level *hlevel = &q->hlevel[cl->level];
382 int prio = ffz(~mask);
383 struct htb_prio *hprio = &hlevel->hprio[prio];
385 mask &= ~(1 << prio);
386 if (hprio->ptr == cl->node + prio)
387 htb_next_rb_node(&hprio->ptr);
389 htb_safe_rb_erase(cl->node + prio, &hprio->row);
390 if (!hprio->row.rb_node)
393 q->row_mask[cl->level] &= ~m;
397 * htb_activate_prios - creates active classe's feed chain
399 * The class is connected to ancestors and/or appropriate rows
400 * for priorities it is participating on. cl->cmode must be new
401 * (activated) mode. It does nothing if cl->prio_activity == 0.
403 static void htb_activate_prios(struct htb_sched *q, struct htb_class *cl)
405 struct htb_class *p = cl->parent;
406 long m, mask = cl->prio_activity;
408 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
414 if (p->un.inner.clprio[prio].feed.rb_node)
415 /* parent already has its feed in use so that
416 * reset bit in mask as parent is already ok
418 mask &= ~(1 << prio);
420 htb_add_to_id_tree(&p->un.inner.clprio[prio].feed, cl, prio);
422 p->prio_activity |= mask;
427 if (cl->cmode == HTB_CAN_SEND && mask)
428 htb_add_class_to_row(q, cl, mask);
432 * htb_deactivate_prios - remove class from feed chain
434 * cl->cmode must represent old mode (before deactivation). It does
435 * nothing if cl->prio_activity == 0. Class is removed from all feed
438 static void htb_deactivate_prios(struct htb_sched *q, struct htb_class *cl)
440 struct htb_class *p = cl->parent;
441 long m, mask = cl->prio_activity;
443 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
450 if (p->un.inner.clprio[prio].ptr == cl->node + prio) {
451 /* we are removing child which is pointed to from
452 * parent feed - forget the pointer but remember
455 p->un.inner.clprio[prio].last_ptr_id = cl->common.classid;
456 p->un.inner.clprio[prio].ptr = NULL;
459 htb_safe_rb_erase(cl->node + prio,
460 &p->un.inner.clprio[prio].feed);
462 if (!p->un.inner.clprio[prio].feed.rb_node)
466 p->prio_activity &= ~mask;
471 if (cl->cmode == HTB_CAN_SEND && mask)
472 htb_remove_class_from_row(q, cl, mask);
475 static inline s64 htb_lowater(const struct htb_class *cl)
478 return cl->cmode != HTB_CANT_SEND ? -cl->cbuffer : 0;
482 static inline s64 htb_hiwater(const struct htb_class *cl)
485 return cl->cmode == HTB_CAN_SEND ? -cl->buffer : 0;
492 * htb_class_mode - computes and returns current class mode
494 * It computes cl's mode at time cl->t_c+diff and returns it. If mode
495 * is not HTB_CAN_SEND then cl->pq_key is updated to time difference
496 * from now to time when cl will change its state.
497 * Also it is worth to note that class mode doesn't change simply
498 * at cl->{c,}tokens == 0 but there can rather be hysteresis of
499 * 0 .. -cl->{c,}buffer range. It is meant to limit number of
500 * mode transitions per time unit. The speed gain is about 1/6.
502 static inline enum htb_cmode
503 htb_class_mode(struct htb_class *cl, s64 *diff)
507 if ((toks = (cl->ctokens + *diff)) < htb_lowater(cl)) {
509 return HTB_CANT_SEND;
512 if ((toks = (cl->tokens + *diff)) >= htb_hiwater(cl))
516 return HTB_MAY_BORROW;
520 * htb_change_class_mode - changes classe's mode
522 * This should be the only way how to change classe's mode under normal
523 * cirsumstances. Routine will update feed lists linkage, change mode
524 * and add class to the wait event queue if appropriate. New mode should
525 * be different from old one and cl->pq_key has to be valid if changing
526 * to mode other than HTB_CAN_SEND (see htb_add_to_wait_tree).
529 htb_change_class_mode(struct htb_sched *q, struct htb_class *cl, s64 *diff)
531 enum htb_cmode new_mode = htb_class_mode(cl, diff);
533 if (new_mode == cl->cmode)
536 if (new_mode == HTB_CANT_SEND)
539 if (cl->prio_activity) { /* not necessary: speed optimization */
540 if (cl->cmode != HTB_CANT_SEND)
541 htb_deactivate_prios(q, cl);
542 cl->cmode = new_mode;
543 if (new_mode != HTB_CANT_SEND)
544 htb_activate_prios(q, cl);
546 cl->cmode = new_mode;
550 * htb_activate - inserts leaf cl into appropriate active feeds
552 * Routine learns (new) priority of leaf and activates feed chain
553 * for the prio. It can be called on already active leaf safely.
554 * It also adds leaf into droplist.
556 static inline void htb_activate(struct htb_sched *q, struct htb_class *cl)
558 WARN_ON(cl->level || !cl->un.leaf.q || !cl->un.leaf.q->q.qlen);
560 if (!cl->prio_activity) {
561 cl->prio_activity = 1 << cl->prio;
562 htb_activate_prios(q, cl);
567 * htb_deactivate - remove leaf cl from active feeds
569 * Make sure that leaf is active. In the other words it can't be called
570 * with non-active leaf. It also removes class from the drop list.
572 static inline void htb_deactivate(struct htb_sched *q, struct htb_class *cl)
574 WARN_ON(!cl->prio_activity);
576 htb_deactivate_prios(q, cl);
577 cl->prio_activity = 0;
580 static void htb_enqueue_tail(struct sk_buff *skb, struct Qdisc *sch,
581 struct qdisc_skb_head *qh)
583 struct sk_buff *last = qh->tail;
596 static int htb_enqueue(struct sk_buff *skb, struct Qdisc *sch,
597 struct sk_buff **to_free)
599 int uninitialized_var(ret);
600 struct htb_sched *q = qdisc_priv(sch);
601 struct htb_class *cl = htb_classify(skb, sch, &ret);
603 if (cl == HTB_DIRECT) {
604 /* enqueue to helper queue */
605 if (q->direct_queue.qlen < q->direct_qlen) {
606 htb_enqueue_tail(skb, sch, &q->direct_queue);
609 return qdisc_drop(skb, sch, to_free);
611 #ifdef CONFIG_NET_CLS_ACT
613 if (ret & __NET_XMIT_BYPASS)
614 qdisc_qstats_drop(sch);
615 __qdisc_drop(skb, to_free);
618 } else if ((ret = qdisc_enqueue(skb, cl->un.leaf.q,
619 to_free)) != NET_XMIT_SUCCESS) {
620 if (net_xmit_drop_count(ret)) {
621 qdisc_qstats_drop(sch);
629 qdisc_qstats_backlog_inc(sch, skb);
631 return NET_XMIT_SUCCESS;
634 static inline void htb_accnt_tokens(struct htb_class *cl, int bytes, s64 diff)
636 s64 toks = diff + cl->tokens;
638 if (toks > cl->buffer)
640 toks -= (s64) psched_l2t_ns(&cl->rate, bytes);
641 if (toks <= -cl->mbuffer)
642 toks = 1 - cl->mbuffer;
647 static inline void htb_accnt_ctokens(struct htb_class *cl, int bytes, s64 diff)
649 s64 toks = diff + cl->ctokens;
651 if (toks > cl->cbuffer)
653 toks -= (s64) psched_l2t_ns(&cl->ceil, bytes);
654 if (toks <= -cl->mbuffer)
655 toks = 1 - cl->mbuffer;
661 * htb_charge_class - charges amount "bytes" to leaf and ancestors
663 * Routine assumes that packet "bytes" long was dequeued from leaf cl
664 * borrowing from "level". It accounts bytes to ceil leaky bucket for
665 * leaf and all ancestors and to rate bucket for ancestors at levels
666 * "level" and higher. It also handles possible change of mode resulting
667 * from the update. Note that mode can also increase here (MAY_BORROW to
668 * CAN_SEND) because we can use more precise clock that event queue here.
669 * In such case we remove class from event queue first.
671 static void htb_charge_class(struct htb_sched *q, struct htb_class *cl,
672 int level, struct sk_buff *skb)
674 int bytes = qdisc_pkt_len(skb);
675 enum htb_cmode old_mode;
679 diff = min_t(s64, q->now - cl->t_c, cl->mbuffer);
680 if (cl->level >= level) {
681 if (cl->level == level)
683 htb_accnt_tokens(cl, bytes, diff);
685 cl->xstats.borrows++;
686 cl->tokens += diff; /* we moved t_c; update tokens */
688 htb_accnt_ctokens(cl, bytes, diff);
691 old_mode = cl->cmode;
693 htb_change_class_mode(q, cl, &diff);
694 if (old_mode != cl->cmode) {
695 if (old_mode != HTB_CAN_SEND)
696 htb_safe_rb_erase(&cl->pq_node, &q->hlevel[cl->level].wait_pq);
697 if (cl->cmode != HTB_CAN_SEND)
698 htb_add_to_wait_tree(q, cl, diff);
701 /* update basic stats except for leaves which are already updated */
703 bstats_update(&cl->bstats, skb);
710 * htb_do_events - make mode changes to classes at the level
712 * Scans event queue for pending events and applies them. Returns time of
713 * next pending event (0 for no event in pq, q->now for too many events).
714 * Note: Applied are events whose have cl->pq_key <= q->now.
716 static s64 htb_do_events(struct htb_sched *q, const int level,
719 /* don't run for longer than 2 jiffies; 2 is used instead of
720 * 1 to simplify things when jiffy is going to be incremented
723 unsigned long stop_at = start + 2;
724 struct rb_root *wait_pq = &q->hlevel[level].wait_pq;
726 while (time_before(jiffies, stop_at)) {
727 struct htb_class *cl;
729 struct rb_node *p = rb_first(wait_pq);
734 cl = rb_entry(p, struct htb_class, pq_node);
735 if (cl->pq_key > q->now)
738 htb_safe_rb_erase(p, wait_pq);
739 diff = min_t(s64, q->now - cl->t_c, cl->mbuffer);
740 htb_change_class_mode(q, cl, &diff);
741 if (cl->cmode != HTB_CAN_SEND)
742 htb_add_to_wait_tree(q, cl, diff);
745 /* too much load - let's continue after a break for scheduling */
746 if (!(q->warned & HTB_WARN_TOOMANYEVENTS)) {
747 pr_warn("htb: too many events!\n");
748 q->warned |= HTB_WARN_TOOMANYEVENTS;
754 /* Returns class->node+prio from id-tree where classe's id is >= id. NULL
755 * is no such one exists.
757 static struct rb_node *htb_id_find_next_upper(int prio, struct rb_node *n,
760 struct rb_node *r = NULL;
762 struct htb_class *cl =
763 rb_entry(n, struct htb_class, node[prio]);
765 if (id > cl->common.classid) {
767 } else if (id < cl->common.classid) {
778 * htb_lookup_leaf - returns next leaf class in DRR order
780 * Find leaf where current feed pointers points to.
782 static struct htb_class *htb_lookup_leaf(struct htb_prio *hprio, const int prio)
786 struct rb_node *root;
787 struct rb_node **pptr;
789 } stk[TC_HTB_MAXDEPTH], *sp = stk;
791 BUG_ON(!hprio->row.rb_node);
792 sp->root = hprio->row.rb_node;
793 sp->pptr = &hprio->ptr;
794 sp->pid = &hprio->last_ptr_id;
796 for (i = 0; i < 65535; i++) {
797 if (!*sp->pptr && *sp->pid) {
798 /* ptr was invalidated but id is valid - try to recover
799 * the original or next ptr
802 htb_id_find_next_upper(prio, sp->root, *sp->pid);
804 *sp->pid = 0; /* ptr is valid now so that remove this hint as it
805 * can become out of date quickly
807 if (!*sp->pptr) { /* we are at right end; rewind & go up */
808 *sp->pptr = sp->root;
809 while ((*sp->pptr)->rb_left)
810 *sp->pptr = (*sp->pptr)->rb_left;
817 htb_next_rb_node(sp->pptr);
820 struct htb_class *cl;
821 struct htb_prio *clp;
823 cl = rb_entry(*sp->pptr, struct htb_class, node[prio]);
826 clp = &cl->un.inner.clprio[prio];
827 (++sp)->root = clp->feed.rb_node;
828 sp->pptr = &clp->ptr;
829 sp->pid = &clp->last_ptr_id;
836 /* dequeues packet at given priority and level; call only if
837 * you are sure that there is active class at prio/level
839 static struct sk_buff *htb_dequeue_tree(struct htb_sched *q, const int prio,
842 struct sk_buff *skb = NULL;
843 struct htb_class *cl, *start;
844 struct htb_level *hlevel = &q->hlevel[level];
845 struct htb_prio *hprio = &hlevel->hprio[prio];
847 /* look initial class up in the row */
848 start = cl = htb_lookup_leaf(hprio, prio);
855 /* class can be empty - it is unlikely but can be true if leaf
856 * qdisc drops packets in enqueue routine or if someone used
857 * graft operation on the leaf since last dequeue;
858 * simply deactivate and skip such class
860 if (unlikely(cl->un.leaf.q->q.qlen == 0)) {
861 struct htb_class *next;
862 htb_deactivate(q, cl);
864 /* row/level might become empty */
865 if ((q->row_mask[level] & (1 << prio)) == 0)
868 next = htb_lookup_leaf(hprio, prio);
870 if (cl == start) /* fix start if we just deleted it */
876 skb = cl->un.leaf.q->dequeue(cl->un.leaf.q);
877 if (likely(skb != NULL))
880 qdisc_warn_nonwc("htb", cl->un.leaf.q);
881 htb_next_rb_node(level ? &cl->parent->un.inner.clprio[prio].ptr:
882 &q->hlevel[0].hprio[prio].ptr);
883 cl = htb_lookup_leaf(hprio, prio);
885 } while (cl != start);
887 if (likely(skb != NULL)) {
888 bstats_update(&cl->bstats, skb);
889 cl->un.leaf.deficit[level] -= qdisc_pkt_len(skb);
890 if (cl->un.leaf.deficit[level] < 0) {
891 cl->un.leaf.deficit[level] += cl->quantum;
892 htb_next_rb_node(level ? &cl->parent->un.inner.clprio[prio].ptr :
893 &q->hlevel[0].hprio[prio].ptr);
895 /* this used to be after charge_class but this constelation
896 * gives us slightly better performance
898 if (!cl->un.leaf.q->q.qlen)
899 htb_deactivate(q, cl);
900 htb_charge_class(q, cl, level, skb);
905 static struct sk_buff *htb_dequeue(struct Qdisc *sch)
908 struct htb_sched *q = qdisc_priv(sch);
911 unsigned long start_at;
913 /* try to dequeue direct packets as high prio (!) to minimize cpu work */
914 skb = __qdisc_dequeue_head(&q->direct_queue);
917 qdisc_bstats_update(sch, skb);
918 qdisc_qstats_backlog_dec(sch, skb);
925 q->now = ktime_get_ns();
928 next_event = q->now + 5LLU * NSEC_PER_SEC;
930 for (level = 0; level < TC_HTB_MAXDEPTH; level++) {
931 /* common case optimization - skip event handler quickly */
933 s64 event = q->near_ev_cache[level];
935 if (q->now >= event) {
936 event = htb_do_events(q, level, start_at);
938 event = q->now + NSEC_PER_SEC;
939 q->near_ev_cache[level] = event;
942 if (next_event > event)
945 m = ~q->row_mask[level];
946 while (m != (int)(-1)) {
950 skb = htb_dequeue_tree(q, prio, level);
951 if (likely(skb != NULL))
955 qdisc_qstats_overlimit(sch);
956 if (likely(next_event > q->now))
957 qdisc_watchdog_schedule_ns(&q->watchdog, next_event);
959 schedule_work(&q->work);
964 /* reset all classes */
965 /* always caled under BH & queue lock */
966 static void htb_reset(struct Qdisc *sch)
968 struct htb_sched *q = qdisc_priv(sch);
969 struct htb_class *cl;
972 for (i = 0; i < q->clhash.hashsize; i++) {
973 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
975 memset(&cl->un.inner, 0, sizeof(cl->un.inner));
978 qdisc_reset(cl->un.leaf.q);
980 cl->prio_activity = 0;
981 cl->cmode = HTB_CAN_SEND;
984 qdisc_watchdog_cancel(&q->watchdog);
985 __qdisc_reset_queue(&q->direct_queue);
987 sch->qstats.backlog = 0;
988 memset(q->hlevel, 0, sizeof(q->hlevel));
989 memset(q->row_mask, 0, sizeof(q->row_mask));
992 static const struct nla_policy htb_policy[TCA_HTB_MAX + 1] = {
993 [TCA_HTB_PARMS] = { .len = sizeof(struct tc_htb_opt) },
994 [TCA_HTB_INIT] = { .len = sizeof(struct tc_htb_glob) },
995 [TCA_HTB_CTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
996 [TCA_HTB_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
997 [TCA_HTB_DIRECT_QLEN] = { .type = NLA_U32 },
998 [TCA_HTB_RATE64] = { .type = NLA_U64 },
999 [TCA_HTB_CEIL64] = { .type = NLA_U64 },
1002 static void htb_work_func(struct work_struct *work)
1004 struct htb_sched *q = container_of(work, struct htb_sched, work);
1005 struct Qdisc *sch = q->watchdog.qdisc;
1008 __netif_schedule(qdisc_root(sch));
1012 static int htb_init(struct Qdisc *sch, struct nlattr *opt,
1013 struct netlink_ext_ack *extack)
1015 struct htb_sched *q = qdisc_priv(sch);
1016 struct nlattr *tb[TCA_HTB_MAX + 1];
1017 struct tc_htb_glob *gopt;
1020 qdisc_watchdog_init(&q->watchdog, sch);
1021 INIT_WORK(&q->work, htb_work_func);
1026 err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
1030 err = nla_parse_nested(tb, TCA_HTB_MAX, opt, htb_policy, NULL);
1034 if (!tb[TCA_HTB_INIT])
1037 gopt = nla_data(tb[TCA_HTB_INIT]);
1038 if (gopt->version != HTB_VER >> 16)
1041 err = qdisc_class_hash_init(&q->clhash);
1045 qdisc_skb_head_init(&q->direct_queue);
1047 if (tb[TCA_HTB_DIRECT_QLEN])
1048 q->direct_qlen = nla_get_u32(tb[TCA_HTB_DIRECT_QLEN]);
1050 q->direct_qlen = qdisc_dev(sch)->tx_queue_len;
1052 if ((q->rate2quantum = gopt->rate2quantum) < 1)
1053 q->rate2quantum = 1;
1054 q->defcls = gopt->defcls;
1059 static int htb_dump(struct Qdisc *sch, struct sk_buff *skb)
1061 struct htb_sched *q = qdisc_priv(sch);
1062 struct nlattr *nest;
1063 struct tc_htb_glob gopt;
1065 /* Its safe to not acquire qdisc lock. As we hold RTNL,
1066 * no change can happen on the qdisc parameters.
1069 gopt.direct_pkts = q->direct_pkts;
1070 gopt.version = HTB_VER;
1071 gopt.rate2quantum = q->rate2quantum;
1072 gopt.defcls = q->defcls;
1075 nest = nla_nest_start(skb, TCA_OPTIONS);
1077 goto nla_put_failure;
1078 if (nla_put(skb, TCA_HTB_INIT, sizeof(gopt), &gopt) ||
1079 nla_put_u32(skb, TCA_HTB_DIRECT_QLEN, q->direct_qlen))
1080 goto nla_put_failure;
1082 return nla_nest_end(skb, nest);
1085 nla_nest_cancel(skb, nest);
1089 static int htb_dump_class(struct Qdisc *sch, unsigned long arg,
1090 struct sk_buff *skb, struct tcmsg *tcm)
1092 struct htb_class *cl = (struct htb_class *)arg;
1093 struct nlattr *nest;
1094 struct tc_htb_opt opt;
1096 /* Its safe to not acquire qdisc lock. As we hold RTNL,
1097 * no change can happen on the class parameters.
1099 tcm->tcm_parent = cl->parent ? cl->parent->common.classid : TC_H_ROOT;
1100 tcm->tcm_handle = cl->common.classid;
1101 if (!cl->level && cl->un.leaf.q)
1102 tcm->tcm_info = cl->un.leaf.q->handle;
1104 nest = nla_nest_start(skb, TCA_OPTIONS);
1106 goto nla_put_failure;
1108 memset(&opt, 0, sizeof(opt));
1110 psched_ratecfg_getrate(&opt.rate, &cl->rate);
1111 opt.buffer = PSCHED_NS2TICKS(cl->buffer);
1112 psched_ratecfg_getrate(&opt.ceil, &cl->ceil);
1113 opt.cbuffer = PSCHED_NS2TICKS(cl->cbuffer);
1114 opt.quantum = cl->quantum;
1115 opt.prio = cl->prio;
1116 opt.level = cl->level;
1117 if (nla_put(skb, TCA_HTB_PARMS, sizeof(opt), &opt))
1118 goto nla_put_failure;
1119 if ((cl->rate.rate_bytes_ps >= (1ULL << 32)) &&
1120 nla_put_u64_64bit(skb, TCA_HTB_RATE64, cl->rate.rate_bytes_ps,
1122 goto nla_put_failure;
1123 if ((cl->ceil.rate_bytes_ps >= (1ULL << 32)) &&
1124 nla_put_u64_64bit(skb, TCA_HTB_CEIL64, cl->ceil.rate_bytes_ps,
1126 goto nla_put_failure;
1128 return nla_nest_end(skb, nest);
1131 nla_nest_cancel(skb, nest);
1136 htb_dump_class_stats(struct Qdisc *sch, unsigned long arg, struct gnet_dump *d)
1138 struct htb_class *cl = (struct htb_class *)arg;
1139 struct gnet_stats_queue qs = {
1141 .overlimits = cl->overlimits,
1145 if (!cl->level && cl->un.leaf.q) {
1146 qlen = cl->un.leaf.q->q.qlen;
1147 qs.backlog = cl->un.leaf.q->qstats.backlog;
1149 cl->xstats.tokens = clamp_t(s64, PSCHED_NS2TICKS(cl->tokens),
1151 cl->xstats.ctokens = clamp_t(s64, PSCHED_NS2TICKS(cl->ctokens),
1154 if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
1155 d, NULL, &cl->bstats) < 0 ||
1156 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1157 gnet_stats_copy_queue(d, NULL, &qs, qlen) < 0)
1160 return gnet_stats_copy_app(d, &cl->xstats, sizeof(cl->xstats));
1163 static int htb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1164 struct Qdisc **old, struct netlink_ext_ack *extack)
1166 struct htb_class *cl = (struct htb_class *)arg;
1171 (new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1172 cl->common.classid, extack)) == NULL)
1175 *old = qdisc_replace(sch, new, &cl->un.leaf.q);
1179 static struct Qdisc *htb_leaf(struct Qdisc *sch, unsigned long arg)
1181 struct htb_class *cl = (struct htb_class *)arg;
1182 return !cl->level ? cl->un.leaf.q : NULL;
1185 static void htb_qlen_notify(struct Qdisc *sch, unsigned long arg)
1187 struct htb_class *cl = (struct htb_class *)arg;
1189 htb_deactivate(qdisc_priv(sch), cl);
1192 static inline int htb_parent_last_child(struct htb_class *cl)
1195 /* the root class */
1197 if (cl->parent->children > 1)
1198 /* not the last child */
1203 static void htb_parent_to_leaf(struct htb_sched *q, struct htb_class *cl,
1204 struct Qdisc *new_q)
1206 struct htb_class *parent = cl->parent;
1208 WARN_ON(cl->level || !cl->un.leaf.q || cl->prio_activity);
1210 if (parent->cmode != HTB_CAN_SEND)
1211 htb_safe_rb_erase(&parent->pq_node,
1212 &q->hlevel[parent->level].wait_pq);
1215 memset(&parent->un.inner, 0, sizeof(parent->un.inner));
1216 parent->un.leaf.q = new_q ? new_q : &noop_qdisc;
1217 parent->tokens = parent->buffer;
1218 parent->ctokens = parent->cbuffer;
1219 parent->t_c = ktime_get_ns();
1220 parent->cmode = HTB_CAN_SEND;
1223 static void htb_destroy_class(struct Qdisc *sch, struct htb_class *cl)
1226 WARN_ON(!cl->un.leaf.q);
1227 qdisc_destroy(cl->un.leaf.q);
1229 gen_kill_estimator(&cl->rate_est);
1230 tcf_block_put(cl->block);
1234 static void htb_destroy(struct Qdisc *sch)
1236 struct htb_sched *q = qdisc_priv(sch);
1237 struct hlist_node *next;
1238 struct htb_class *cl;
1241 cancel_work_sync(&q->work);
1242 qdisc_watchdog_cancel(&q->watchdog);
1243 /* This line used to be after htb_destroy_class call below
1244 * and surprisingly it worked in 2.4. But it must precede it
1245 * because filter need its target class alive to be able to call
1246 * unbind_filter on it (without Oops).
1248 tcf_block_put(q->block);
1250 for (i = 0; i < q->clhash.hashsize; i++) {
1251 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1252 tcf_block_put(cl->block);
1256 for (i = 0; i < q->clhash.hashsize; i++) {
1257 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1259 htb_destroy_class(sch, cl);
1261 qdisc_class_hash_destroy(&q->clhash);
1262 __qdisc_reset_queue(&q->direct_queue);
1265 static int htb_delete(struct Qdisc *sch, unsigned long arg)
1267 struct htb_sched *q = qdisc_priv(sch);
1268 struct htb_class *cl = (struct htb_class *)arg;
1269 struct Qdisc *new_q = NULL;
1272 /* TODO: why don't allow to delete subtree ? references ? does
1273 * tc subsys guarantee us that in htb_destroy it holds no class
1274 * refs so that we can remove children safely there ?
1276 if (cl->children || cl->filter_cnt)
1279 if (!cl->level && htb_parent_last_child(cl)) {
1280 new_q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1281 cl->parent->common.classid,
1289 unsigned int qlen = cl->un.leaf.q->q.qlen;
1290 unsigned int backlog = cl->un.leaf.q->qstats.backlog;
1292 qdisc_reset(cl->un.leaf.q);
1293 qdisc_tree_reduce_backlog(cl->un.leaf.q, qlen, backlog);
1296 /* delete from hash and active; remainder in destroy_class */
1297 qdisc_class_hash_remove(&q->clhash, &cl->common);
1299 cl->parent->children--;
1301 if (cl->prio_activity)
1302 htb_deactivate(q, cl);
1304 if (cl->cmode != HTB_CAN_SEND)
1305 htb_safe_rb_erase(&cl->pq_node,
1306 &q->hlevel[cl->level].wait_pq);
1309 htb_parent_to_leaf(q, cl, new_q);
1311 sch_tree_unlock(sch);
1313 htb_destroy_class(sch, cl);
1317 static int htb_change_class(struct Qdisc *sch, u32 classid,
1318 u32 parentid, struct nlattr **tca,
1319 unsigned long *arg, struct netlink_ext_ack *extack)
1322 struct htb_sched *q = qdisc_priv(sch);
1323 struct htb_class *cl = (struct htb_class *)*arg, *parent;
1324 struct nlattr *opt = tca[TCA_OPTIONS];
1325 struct nlattr *tb[TCA_HTB_MAX + 1];
1326 struct tc_htb_opt *hopt;
1330 /* extract all subattrs from opt attr */
1334 err = nla_parse_nested(tb, TCA_HTB_MAX, opt, htb_policy, NULL);
1339 if (tb[TCA_HTB_PARMS] == NULL)
1342 parent = parentid == TC_H_ROOT ? NULL : htb_find(parentid, sch);
1344 hopt = nla_data(tb[TCA_HTB_PARMS]);
1345 if (!hopt->rate.rate || !hopt->ceil.rate)
1348 /* Keeping backward compatible with rate_table based iproute2 tc */
1349 if (hopt->rate.linklayer == TC_LINKLAYER_UNAWARE)
1350 qdisc_put_rtab(qdisc_get_rtab(&hopt->rate, tb[TCA_HTB_RTAB],
1353 if (hopt->ceil.linklayer == TC_LINKLAYER_UNAWARE)
1354 qdisc_put_rtab(qdisc_get_rtab(&hopt->ceil, tb[TCA_HTB_CTAB],
1357 if (!cl) { /* new class */
1358 struct Qdisc *new_q;
1362 struct gnet_estimator opt;
1365 .nla_len = nla_attr_size(sizeof(est.opt)),
1366 .nla_type = TCA_RATE,
1369 /* 4s interval, 16s averaging constant */
1375 /* check for valid classid */
1376 if (!classid || TC_H_MAJ(classid ^ sch->handle) ||
1377 htb_find(classid, sch))
1380 /* check maximal depth */
1381 if (parent && parent->parent && parent->parent->level < 2) {
1382 pr_err("htb: tree is too deep\n");
1386 cl = kzalloc(sizeof(*cl), GFP_KERNEL);
1390 err = tcf_block_get(&cl->block, &cl->filter_list, sch, extack);
1395 if (htb_rate_est || tca[TCA_RATE]) {
1396 err = gen_new_estimator(&cl->bstats, NULL,
1399 qdisc_root_sleeping_running(sch),
1400 tca[TCA_RATE] ? : &est.nla);
1402 tcf_block_put(cl->block);
1409 RB_CLEAR_NODE(&cl->pq_node);
1411 for (prio = 0; prio < TC_HTB_NUMPRIO; prio++)
1412 RB_CLEAR_NODE(&cl->node[prio]);
1414 /* create leaf qdisc early because it uses kmalloc(GFP_KERNEL)
1415 * so that can't be used inside of sch_tree_lock
1416 * -- thanks to Karlis Peisenieks
1418 new_q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1421 if (parent && !parent->level) {
1422 unsigned int qlen = parent->un.leaf.q->q.qlen;
1423 unsigned int backlog = parent->un.leaf.q->qstats.backlog;
1425 /* turn parent into inner node */
1426 qdisc_reset(parent->un.leaf.q);
1427 qdisc_tree_reduce_backlog(parent->un.leaf.q, qlen, backlog);
1428 qdisc_destroy(parent->un.leaf.q);
1429 if (parent->prio_activity)
1430 htb_deactivate(q, parent);
1432 /* remove from evt list because of level change */
1433 if (parent->cmode != HTB_CAN_SEND) {
1434 htb_safe_rb_erase(&parent->pq_node, &q->hlevel[0].wait_pq);
1435 parent->cmode = HTB_CAN_SEND;
1437 parent->level = (parent->parent ? parent->parent->level
1438 : TC_HTB_MAXDEPTH) - 1;
1439 memset(&parent->un.inner, 0, sizeof(parent->un.inner));
1441 /* leaf (we) needs elementary qdisc */
1442 cl->un.leaf.q = new_q ? new_q : &noop_qdisc;
1444 cl->common.classid = classid;
1445 cl->parent = parent;
1447 /* set class to be in HTB_CAN_SEND state */
1448 cl->tokens = PSCHED_TICKS2NS(hopt->buffer);
1449 cl->ctokens = PSCHED_TICKS2NS(hopt->cbuffer);
1450 cl->mbuffer = 60ULL * NSEC_PER_SEC; /* 1min */
1451 cl->t_c = ktime_get_ns();
1452 cl->cmode = HTB_CAN_SEND;
1454 /* attach to the hash list and parent's family */
1455 qdisc_class_hash_insert(&q->clhash, &cl->common);
1458 if (cl->un.leaf.q != &noop_qdisc)
1459 qdisc_hash_add(cl->un.leaf.q, true);
1461 if (tca[TCA_RATE]) {
1462 err = gen_replace_estimator(&cl->bstats, NULL,
1465 qdisc_root_sleeping_running(sch),
1473 rate64 = tb[TCA_HTB_RATE64] ? nla_get_u64(tb[TCA_HTB_RATE64]) : 0;
1475 ceil64 = tb[TCA_HTB_CEIL64] ? nla_get_u64(tb[TCA_HTB_CEIL64]) : 0;
1477 psched_ratecfg_precompute(&cl->rate, &hopt->rate, rate64);
1478 psched_ratecfg_precompute(&cl->ceil, &hopt->ceil, ceil64);
1480 /* it used to be a nasty bug here, we have to check that node
1481 * is really leaf before changing cl->un.leaf !
1484 u64 quantum = cl->rate.rate_bytes_ps;
1486 do_div(quantum, q->rate2quantum);
1487 cl->quantum = min_t(u64, quantum, INT_MAX);
1489 if (!hopt->quantum && cl->quantum < 1000) {
1493 if (!hopt->quantum && cl->quantum > 200000) {
1495 cl->quantum = 200000;
1498 cl->quantum = hopt->quantum;
1499 if ((cl->prio = hopt->prio) >= TC_HTB_NUMPRIO)
1500 cl->prio = TC_HTB_NUMPRIO - 1;
1503 cl->buffer = PSCHED_TICKS2NS(hopt->buffer);
1504 cl->cbuffer = PSCHED_TICKS2NS(hopt->cbuffer);
1506 sch_tree_unlock(sch);
1509 pr_warn("HTB: quantum of class %X is %s. Consider r2q change.\n",
1510 cl->common.classid, (warn == -1 ? "small" : "big"));
1512 qdisc_class_hash_grow(sch, &q->clhash);
1514 *arg = (unsigned long)cl;
1521 static struct tcf_block *htb_tcf_block(struct Qdisc *sch, unsigned long arg,
1522 struct netlink_ext_ack *extack)
1524 struct htb_sched *q = qdisc_priv(sch);
1525 struct htb_class *cl = (struct htb_class *)arg;
1527 return cl ? cl->block : q->block;
1530 static unsigned long htb_bind_filter(struct Qdisc *sch, unsigned long parent,
1533 struct htb_class *cl = htb_find(classid, sch);
1535 /*if (cl && !cl->level) return 0;
1536 * The line above used to be there to prevent attaching filters to
1537 * leaves. But at least tc_index filter uses this just to get class
1538 * for other reasons so that we have to allow for it.
1540 * 19.6.2002 As Werner explained it is ok - bind filter is just
1541 * another way to "lock" the class - unlike "get" this lock can
1542 * be broken by class during destroy IIUC.
1546 return (unsigned long)cl;
1549 static void htb_unbind_filter(struct Qdisc *sch, unsigned long arg)
1551 struct htb_class *cl = (struct htb_class *)arg;
1557 static void htb_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1559 struct htb_sched *q = qdisc_priv(sch);
1560 struct htb_class *cl;
1566 for (i = 0; i < q->clhash.hashsize; i++) {
1567 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1568 if (arg->count < arg->skip) {
1572 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1581 static const struct Qdisc_class_ops htb_class_ops = {
1584 .qlen_notify = htb_qlen_notify,
1586 .change = htb_change_class,
1587 .delete = htb_delete,
1589 .tcf_block = htb_tcf_block,
1590 .bind_tcf = htb_bind_filter,
1591 .unbind_tcf = htb_unbind_filter,
1592 .dump = htb_dump_class,
1593 .dump_stats = htb_dump_class_stats,
1596 static struct Qdisc_ops htb_qdisc_ops __read_mostly = {
1597 .cl_ops = &htb_class_ops,
1599 .priv_size = sizeof(struct htb_sched),
1600 .enqueue = htb_enqueue,
1601 .dequeue = htb_dequeue,
1602 .peek = qdisc_peek_dequeued,
1605 .destroy = htb_destroy,
1607 .owner = THIS_MODULE,
1610 static int __init htb_module_init(void)
1612 return register_qdisc(&htb_qdisc_ops);
1614 static void __exit htb_module_exit(void)
1616 unregister_qdisc(&htb_qdisc_ops);
1619 module_init(htb_module_init)
1620 module_exit(htb_module_exit)
1621 MODULE_LICENSE("GPL");